When it comes to artificial intelligence chips, NVIDIA currently has a considerable lead… So much so, that the company finds itself in a monopoly situation with its A100, H100 and other chips. However, things could change dramatically with Intel’s Gaudi 2, a rocketship made by Intel.
Gaudi 2: NVIDIA’s monopoly called into question?
Intel recently published the results of its Gaudi 2 chips in the Artificial Intelligence sector. Using FP8 instructions, the brand has managed to drastically improve the performance of its chips. As can be seen, the gains are as high as 103%, compared with initial projections of 90%. In the benchmark, on MLPerf Training 3.1 GPT-3, training of the 384 cards was completed in 153.58 minutes, compared with 311.94 minutes in June.
Now, in the face of NVIDIA’s offering, the gap between the cards is very narrow at 9% with the H100. Against the GH200-96G, the gap is a little wider in NVIDIA’s favor: 12%, all measured on GPTJ-99 and 8 gas pedals. Well, that’s online, since offline, NVIDIA still boasts 28% higher performance with an H100. With the GH200, this difference is reduced to 27%. However, it’s worth noting that the A100 is dropped every time.
Now for a small but important fact: the 99.9% calculation accuracy in FP8.
In short, all these benchmarks are very encouraging for Intel. All that remains to be seen is the price of such a solution. As a reminder, an NVIDIA H100 costs at least $30,000, if not more, with Amazon listing it at $43,989. Whatever the case, Intel wants to position itself as a competitive price/performance ratio in the sector!